Deep Learning Based SAR Image Super Resolution

High-resolution synthetic aperture radar(SAR)images have important application value in many fields.However,due to unfavorable factors such as hardware performance limitations,complex imaging conditions,and the inherent multiplicative speckle noise,the resolution of the SAR image obtained has been severely reduced.The image super-resolution reconstruction,which known as a method that can not only improve the resolution of a image,but also introduce new high-frequency information,such as edges and textures,is a efficient way to eliminate the noise and clutter and further enhance the image quality.In recent years,the super-resolution of natural images based on deep learning has achieved great success,but there are very few researches on the superresolution of SAR images.Therefore,this thesis applies deep learning technology to SAR image super-resolution reconstruction,using the most popular convolutional neural network(CNN)and generative countermeasure network(GAN)models,and introducing a variety of advanced structures and mechanisms in deep learning.Finally,it is possible to convert low-resolution SAR images into high-quality,high-resolution SAR images.The specific research content of the thesis is as follows:1)A super-resolution reconstruction network for SAR images based on convolutional neural network(CNN)is proposed.First,the network uses three different scale convolution kernels to extract image features to ensure the diversity of features.At the same time,it combines the global residual structure(GRB)in the very deep super resolution(VDSR)network and the local residual structure(LRB)in Res Net.This helps the network only learn a small part of non-zero pixel values,which speeds up network convergence and reduces learning difficulty.As for the selection of the up-sampling layer,the thesis discarded the conventional transposed convolution or sub-pixel convolution layer,but adopted a more efficient way of increasing the number of convolution kernel channels and then rearranging pixels.Moreover,the up-sampling operation is performed at the back end of the network to ensure the network feature learning process mostly in low-resolution space,which can effectively reduce the training time of the network.In addition,the dense connection structure is also introduced into the network,so that the network could make full use of the feature information of the image,which is more conducive to the recovery of high-frequency information of the image.The experimental results show that the SAR image reconstructed by the network has achieved significant improvement in aspect of peak signal-to-noise ratio,structural similarity and edge retention when compared with the traditional super-resolution method.Even compared with the state-of-the-art deep-learning based method,there also be an improvement.When it comes to ship detection in largescene SAR images,the detection accuracy of images which have been super-resolved will be significantly improved.There is also a certain degree of improvement when compared to other super resolution methods.2)A SAR image super-resolution reconstruction network based on generative adversarial networks(GAN)is proposed.In this thesis,the generator and discriminator are improved on the basis of the super-resolution generative adversarial network(SRGAN).For generator,the Laplacian pyramid structure is used to magnify the image step by step to achieve larger image super-resolution reconstruction.Moreover,channel attention mechanism is introduced to further enhance the useful features.It will eliminate the influence of coherent SAR image speckle noise to a certain extent.The discriminator uses a novel relative discriminator structure,so that the real image and the generated image can be used to optimize the network simultaneously.The loss function of the network is composed of three parts: pixel-wise mean square error(MSE)loss,adversarial loss and perceptual loss.The perceptual loss is based on the high-level abstract features extracted by the VGG19 network,which is helpful to the recovery of high-frequency information such as the edge texture of the SAR image.In the end,the network proposed in this thesis can reconstruct very realistic,high-resolution SAR images with clear texture.And in ship inspection task,the detection accuracy has been greatly improved compared with low-resolution images.It is also slightly better than other super-resolution methods.